Transformer-based large language models (LLMs) are the foundation of the modern generative AI landscape.

Transformers aren’t the only way to do gen AI, though. Over the course of the last year, Mamba, an approach that uses Structured State Space Models (SSM), has also picked up adoption as an alternative approach from multiple vendors, including AI21 and AI silicon giant Nvidia. 

Nvidia first discussed the concept of Mamba-powered models in 2024 when it initially released the MambaVision research and some early models. This week, Nvidia is expanding on its initial effort with a series of updated MambaVision models available on Hugging Face.

MambaVision, as the name implies, is a Mamba-based model family for computer vision and image recognition tasks. The promise of MambaVision for enterprise is that it could improve the efficiency and accuracy of vision operations, at potentially lower costs, thanks to lower computational requirements.

What are SSMs and how do they compare to transformers?

SSMs are a neural network architecture class that processes sequential data differently from traditional transformers. 

While transformers use attention mechanisms to process all tokens in relation to each other, SSMs model sequence data as a continuous dynamic system.

Mamba is a specific SSM implementation developed to address the limitations of earlier SSM models. It introduces selective state space modelling that dynamically adapts to input data and hardware-aware design for efficient GPU utilization. Mamba aims to provide comparable performance to transformers on many tasks while using fewer computational resources

Nvidia using hybrid architecture with MambaVision to revolutionize Computer Vision

Traditional Vision Transformers (ViT) have dominated high-performance computer vision for the last several years, but at significant computational cost. Pure Mamba-based approaches, while more efficient, have struggled to match Transformer performance on complex vision tasks requiring global context understanding.

MambaVision bridges this gap by adopting a hybrid approach. Nvidia’s MambaVision is a hybrid model that strategically combines Mamba’s efficiency with the Transformer’s modelling power. 

The architecture’s innovation lies in its redesigned Mamba formulation specifically engineered for visual feature modeling, augmented by strategic placement of self-attention blocks in the final layers to capture complex spatial dependencies.

Unlike conventional vision models that rely exclusively on either attention mechanisms or convolutional approaches, MambaVision’s hierarchical architecture employs both paradigms simultaneously. The model processes visual information through sequential scan-based operations from Mamba while leveraging self-attention to model global context — effectively getting the best of both worlds.

MambaVision now has 740 million parameters

The new set of MambaVision models released on Hugging Face is available under the Nvidia Source Code License-NC, which is an open license.

The initial variants of MambaVision released in 2024 include the T and T2 variants, which were trained on the ImageNet-1K library. The new models released this week include the L/L2 and L3 variants, which are scaled-up models.

“Since the initial release, we’ve significantly enhanced MambaVision, scaling it up to an impressive 740 million parameters,” Ali Hatamizadeh, Senior Research Scientist at Nvidia wrote in a Hugging Face discussion post. “We’ve also expanded our training approach by utilizing the larger ImageNet-21K dataset and have introduced native support for higher resolutions, now handling images at 256 and 512 pixels compared to the original 224 pixels.”

According to Nvidia, the improved scale in the new MambaVision models also improves performance.

Independent AI consultant Alex Fazio explained to VentureBeat that the new MambaVision models’ training on larger datasets makes them much better at handling more diverse and complex tasks. 

He noted that the new models include high-resolution variants perfect for detailed image analysis. Fazio said that the lineup has also expanded with advanced configurations offering more flexibility and scalability for different workloads.

“In terms of benchmarks, the 2025 models are expected to outperform the 2024 ones because they generalize better across larger datasets and tasks, Fazio said. 

Enterprise implications of MambaVision

For enterprises building computer vision applications, MambaVision’s balance of performance and efficiency opens new possibilities

Reduced inference costs: The improved throughput means lower GPU compute requirements for similar performance levels compared to Transformer-only models.

Edge deployment potential: While still large, MambaVision’s architecture is more amenable to optimization for edge devices than pure Transformer approaches.

Improved downstream task performance: The gains on complex tasks like object detection and segmentation translate directly to better performance for real-world applications like inventory management, quality control, and autonomous systems.

Simplified deployment: NVIDIA has released MambaVision with Hugging Face integration, making implementation straightforward with just a few lines of code for both classification and feature extraction.

What this means for enterprise AI strategy

MambaVision represents an opportunity for enterprises to deploy more efficient computer vision systems that maintain high accuracy. The model’s strong performance means that it can potentially serve as a versatile foundation for multiple computer vision applications across industries.

MambaVision is still somewhat of an early effort, but it does represent a glimpse into the future of computer vision models.

MambaVision highlights how architectural innovation—not just scale—continues to drive meaningful improvements in AI capabilities. Understanding these architectural advances is becoming increasingly crucial for technical decision-makers to make informed AI deployment choices.